AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Efficient, Property-Aligned Fan-Out Retrieval via RL-Compiled Diffusion

Pengcheng Jiang, Judith Yue Li, Moonkyung Ryu, Lily Hu, Kun Su, Zhong Yi Wan, Liam Hebert, Hao Peng, Jiawei Han, Dima Kuzmin, Craig Boutilier

ICML 2026 · Reinforcement learning › RL from feedback · approach: GRPO stability and variants · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
RL-train the fan-out LM with set-level rewardsGemma3-4B fan-out · Polyvore OARaverage score: 38.5 -> 49.1Improved
RL-train the fan-out LM with set-level rewardsQwen3-4B fan-out · Polyvore OARaverage score: 28.1 -> 42.6Improved
RL-train the fan-out LM instead of Best-of-N samplingGemma3-4B · Music OARaverage score: 49.2 -> 58.1Improved
use zero-shot fan-out instead of no fan-outGemini-2.5-Flash · Polyvore OARaverage score: 26.1 -> 31.5Improved
Best-of-N selection with N=5 using training rewardsGemma3-4B zero-shot · Polyvore OARaverage score: 38.5 -> 40.9Improved
distill FOLM into diffusion retrieverR4T · Polyvore OAR (Gemma3-4B)diversity: 76.8 -> 74.3No clear effect
distill FOLM into diffusion retrieverR4T · Polyvore WSCR (Gemma3-4B)Vendi Score: 40.5 -> 46.2Improved
distill FOLM into diffusion retrieverR4T · Polyvore WSCR (Gemma3-4B)Recall@5K: 16.9 -> 15.0Hurt
replace autoregressive fan-out with diffusion single passfan-out generation · query fan-out latencylatency: 1.46 s -> 0.07 s at batch 8Improved
train with groundedness reward only (no diversity) (ablation)FOLM RL · Polyvore OARtraining stabilityHurt
train with groundedness and alignment, no diversity (ablation)FOLM RL · Polyvore OARtraining stabilityHurt
add diversity reward to groundedness and alignment (ablation)FOLM RL · Polyvore OARtraining stabilityImproved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy